Merge branch 'comfyanonymous:master' into master

Author: kinqsradio

.github/workflows/stale-issues.yml

@@ -0,0 +1,21 @@
+name: 'Close stale issues'
+on:
+  schedule:
+    # Run daily at 430 am PT
+    - cron: '30 11 * * *'
+permissions:
+  issues: write
+
+jobs:
+  stale:
+    runs-on: ubuntu-latest
+    steps:
+      - uses: actions/stale@v9
+        with:
+          stale-issue-message: "This issue is being marked stale because it has not had any activity for 30 days. Reply below within 7 days if your issue still isn't solved, and it will be left open. Otherwise, the issue will be closed automatically."
+          days-before-stale: 30
+          days-before-close: 7
+          stale-issue-label: 'Stale'
+          only-labels: 'User Support'
+          exempt-all-assignees: true
+          exempt-all-milestones: true

README.md

@@ -135,17 +135,17 @@ Put your VAE in: models/vae
 ### AMD GPUs (Linux only)
 AMD users can install rocm and pytorch with pip if you don't have it already installed, this is the command to install the stable version:
 
-```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.0```
+```pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/rocm6.1```
 
-This is the command to install the nightly with ROCm 6.0 which might have some performance improvements:
+This is the command to install the nightly with ROCm 6.2 which might have some performance improvements:
 
-```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.1```
+```pip install --pre torch torchvision torchaudio --index-url https://download.pytorch.org/whl/nightly/rocm6.2```
 
 ### NVIDIA
 
 Nvidia users should install stable pytorch using this command:
 
-```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu121```
+```pip install torch torchvision torchaudio --extra-index-url https://download.pytorch.org/whl/cu124```
 
 This is the command to install pytorch nightly instead which might have performance improvements:
 

comfy/model_management.py

@@ -370,12 +370,11 @@ def offloaded_memory(loaded_models, device):
             offloaded_mem += m.model_offloaded_memory()
     return offloaded_mem
 
-def minimum_inference_memory():
-    return (1024 * 1024 * 1024) * 1.2
+WINDOWS = any(platform.win32_ver())
 
-EXTRA_RESERVED_VRAM = 200 * 1024 * 1024
-if any(platform.win32_ver()):
-    EXTRA_RESERVED_VRAM = 500 * 1024 * 1024 #Windows is higher because of the shared vram issue
+EXTRA_RESERVED_VRAM = 400 * 1024 * 1024
+if WINDOWS:
+    EXTRA_RESERVED_VRAM = 600 * 1024 * 1024 #Windows is higher because of the shared vram issue
 
 if args.reserve_vram is not None:
     EXTRA_RESERVED_VRAM = args.reserve_vram * 1024 * 1024 * 1024
@@ -384,6 +383,9 @@ def minimum_inference_memory():
 def extra_reserved_memory():
     return EXTRA_RESERVED_VRAM
 
+def minimum_inference_memory():
+    return (1024 * 1024 * 1024) * 0.8 + extra_reserved_memory()
+
 def unload_model_clones(model, unload_weights_only=True, force_unload=True):
     to_unload = []
     for i in range(len(current_loaded_models)):
@@ -1002,7 +1004,10 @@ def should_use_fp16(device=None, model_params=0, prioritize_performance=True, ma
     nvidia_10_series = ["1080", "1070", "titan x", "p3000", "p3200", "p4000", "p4200", "p5000", "p5200", "p6000", "1060", "1050", "p40", "p100", "p6", "p4"]
     for x in nvidia_10_series:
         if x in props.name.lower():
-            return True
+            if WINDOWS or manual_cast:
+                return True
+            else:
+                return False #weird linux behavior where fp32 is faster
 
     if manual_cast:
         free_model_memory = maximum_vram_for_weights(device)